Optimization of accelerated solvent extraction of zeaxanthin from orange paprika using response surface methodology and an artificial neural network coupled with a genetic algorithm

Abstract

This study aimed to optimize the accelerated solvent extraction (ASE) condition of zeaxanthin from orange paprika using a response surface methodology (RSM) or an artificial neural network (ANN) with a genetic algorithm (GA). Input variables were ethanol concentration, extraction time, and extraction temperature, while output variable was zeaxanthin. The mean squared error and regression correlation coefficient of the developed ANN model were 0.3038 and 0.9983, respectively. Predicted optimal extraction conditions from ANN-GA for maximum zeaxanthin were 100% ethanol, 3.4 min, and 99.2 °C. The relative errors under the optimal extraction conditions were RSM for 10.46% and ANN-GA for 2.18%. We showed that the recovery of hydrophobic zeaxanthin could be performed using ethanol, an eco-friendly solvent, via ASE, and the extraction efficiency could be improved by ANN-GA modeling than RSM. Therefore, combining ASE and ANN-GA might be desirable for the efficient and eco-friendly extraction of hydrophobic functional materials from food resources.

Supplementary Information

The online version contains supplementary material available at 10.1007/s10068-023-01514-8.

Introduction

Zeaxanthin is a major carotenoid in plants synthesized from lycopene or beta-carotene and can change to antheraxanthin according to light exposure in the xanthophyll cycles (Demmig-Adams et al., 2020). Zeaxanthin has a 3-hydroxy-β-end group and is a stereoisomer of lutein, a carotenoid with eye-health-promoting effects (Murillo et al., 2019; Maeda et al., 2021). Zeaxanthin has several health-promoting effects, such as promoting eye health, reducing age-related macular degeneration, and exerting anticancer effects (Bouyahya et al., 2021; Mrowicka et al., 2022). Zeaxanthin is mainly contained in plant foods such as dark green leafy vegetables (kale and spinach), paprika, corn, saffron, and wolfberries. Among them, paprika (Capsicum annuum L.) is a major source of various carotenoids, including zeaxanthin. Paprika exhibits diverse varieties, each characterized by distinct colors and shapes. Additionally, paprika contains many bioactive compounds, including carotenoids, tocopherols, ascorbic acid, and flavonoids (Nazzaro et al., 2009). Among them, orange paprika has significantly higher zeaxanthin than other paprika colors (Kye et al., 2022). Therefore, supplementing orange paprika with zeaxanthin is expected to have health-promoting effects. Developing a suitable and efficient extraction method is crucial for effectively utilizing zeaxanthin from plant sources like orange paprika as a functional food ingredient. Currently, zeaxanthin extraction methods mainly use many organic solvents, such as acetone, n-hexane, toluene, and petroleum ether, or supercritical fluid extraction; however, there are some limitations, such as low yield and high cost (Shen et al., 2022). Furthermore, the excessive organic solvents used for extracting and recovering functional materials might cause environmental hazards, especially water pollution and soil contamination.

Accelerated solvent extraction (ASE) is an automatic natural compound extraction technique with higher efficiency than conventional extraction techniques. In contrast to conventional extraction methods, ASE needs a dedicated machine. ASE can be conducted under intense pressure conditions by placing the target material within a metal cell of exceptionally high strength. Based on these characteristics, ASE consumes small amounts of solvents and has shorter extraction times than other conventional extraction techniques because it can be performed under high pressure (Al Jitan et al., 2018; Kim et al., 2019). The extraction solvents used in conventional extraction methods and ASE are generally similar; however, while traditional methods involve a manual process, ASE utilizes an automated high-pressure system. This extraction method induces tissue to collapse and improves extraction efficiency. Therefore, when extracting natural products, especially carotenoids, from vegetables and fruits, ASE can be more efficient than conventional extraction methods (Sun et al., 2012; Lee et al., 2022). ASE extracts at a relatively low temperature compared to traditional extraction methods, so it can be advantageous for extracting substances that are easily denatured by heat (Al Jitan et al., 2018). Controlling extraction variables such as extraction temperature, pressure, and solvent are essential to utilize ASE successfully (Quintana et al., 2007; Lee et al., 2022).

Response surface methodology (RSM) is a common approach for optimization in food science fields. RSM allows for studying the effects of multiple factors and is a valuable mathematical and statistical method for optimizing independent factors in a range of experiments. RSM encompasses a sophisticated set of statistical and mathematical techniques to optimize processes and models by determining the interaction effects of independent variables (Ciric et al., 2020; Lee et al., 2022).

An artificial neural network (ANN) is a deep learning method that imitates the structure and function of the human nervous system (Lee et al., 2022). It is suitable for modeling complex nonlinear relationships between multiple variables and has shown to be more accurate than other fitting techniques, such as RSM, in generating predictive models (Ciric et al., 2020). On the other hand, a genetic algorithm (GA) is a potent optimization method that draws inspiration from the evolutionary mechanisms found in biological systems. These mechanisms involve processes such as selection, crossover, and mutation (Mirjalili, 2019). The significant advantage of the GA-based method is its capacity to find a simplified solution for extremely complex nonlinear problems. Therefore, it has been attempted for many applications in natural compound research areas, such as optimizing extraction methods for most target compounds. Using an ANN coupled with a GA maximizes the advantages of both tools. The ANN-GA method can help accurately find the optimal extraction conditions with much less effort than conventional optimization methods (Zhang et al., 2020). Recently, it has been employed to optimize extraction conditions for carotenoids from a wide range of food sources. (Sarkar et al., 2020; Patra et al., 2021; Sharayei et al., 2021).

This study aimed to utilize and compare an RSM and ANN-GA to identify the optimal extract conditions, including ethanol concentration, extraction time, and extraction temperature, for maximizing the yield of zeaxanthin extracted from orange paprika using ASE.

Materials and methods

Chemicals

Acetonitrile and methanol were purchased from J.T. Baker (Phillipsburg, NJ, USA). Ethanol was purchased from Samchun Pure Chemicals (Pyeongtaek, Korea). Lutein, α-cryptoxanthin, β-cryptoxanthin, β-carotene, and zeaxanthin were purchased from Carotenature (Lupsingen, Switzerland). Diethyl ether, sodium chloride, sodium sulfate, and potassium hydroxide were purchased from Junsei Chemical Co. Ltd. (Tokyo, Japan).

Sample preparation

Orange paprika, harvested in Jinju, Korea, was kindly provided by Gyeongsangnam-do Agricultural Research & Extension Services (Jinju, Korea). The paprika was washed off with tap water, chopped using a knife after removing the stalk, seed, and placenta, freeze-dried for three days, and finely ground to pass through a 35 mesh test sieve (Chunggye Sanggongsa, Seoul, Korea). The powder was stored at − 20 °C until extraction or analysis.

Extraction of zeaxanthin

Zeaxanthin extraction was performed by an accelerated solvent extractor (ASE 350, Dionex, Sunnyvale, CA, USA) according to a previous study (Kim et al., 2016) (Fig. S1). Briefly, the paprika powder (1 g) was mixed with ASE Prep diatomaceous earth and loaded into an extraction cell with a cellulose filter (Whatman Co., Maidstone, UK). ASE conditions were 1500 psi pressure, 60 s nitrogen purge, and 3 static cycles. Various ethanol concentrations (50–100%), extraction times (3–9 min), and temperatures (60–100 °C) were used to optimize zeaxanthin extraction. The extract was concentrated and dried using a centrifugal vacuum concentrator (Labogene, Seoul, Korea).

Saponification of the extract

Saponification of the extract for carotenoid analysis was performed according to a previous study with some modifications (Kim et al., 2017). Briefly, the dried extract was dissolved in 2 mL of ethanol (extract solvent) and mixed with methanol (3 mL) and 30% potassium hydroxide in methanol (1 mL). After extraction with diethyl ether (20 mL), the extract was washed with distilled water, added 10% sodium chloride (5 mL) and 2% sodium sulfate (5 mL), separated into hydrophilic and hydrophobic phases. The hydrophobic phase was collected, evaporated, and dissolved in acetone (1 mL) for storage at – 70 ℃ for further analysis.

UPLC analysis of zeaxanthin and other carotenoids

Analysis of zeaxanthin and other carotenoids was performed according to previous studies (Kim et al., 2019; Kye et al., 2022). Zeaxanthin and other carotenoids were determined using an ACQUITY UPLC H-Class system (Waters, Milford, MA, USA) coupled with a TUV detector (Waters) equipped with an ACQUITY HSS T3 column (100 mm × 2.1 mm, 1.8 μm; Waters) heated at 35 °C. Mobile phases A and B were distilled water and acetonitrile/methanol/dichloromethane (65:25:10, v:v:v), respectively. Gradient elution was set as follows: 70% B in 0–6.5 min, 70–75% B in 6.5–7 min, 75% B in 7–11 min, 75–70% B in 11–11.5 min, 70% B in 11.5–17 min, 70–100% B in 17–17.5 min, 100% B in 17.5–27.5 min, 100–70% B in 27.5–28 min, and 70% B in 28–30 min. The injection volume, flow rate, and detection wavelength were 1 μL, 0.5 mL/min, and 450 nm, respectively.

Experimental design for RSM and ANN modeling to optimize extract conditions of zeaxanthin in orange paprika

For the experimental design, ethanol concentration (X₁, 50–100%), extraction time (X₂, 3–9 min), and extraction temperature (X₃, 60–100 °C) were selected as independent input variables, and their various combinations were obtained using central composite design (CCD) as per Table 1. Planning of experimental runs was done using Design-Expert 13 (Stat-Ease, MN, USA). A total of 20 experimental runs were conducted, and zeaxanthin content (Y) was calculated for each run (Table 1). RSM and ANN modeling was then used to analyze the experimental design.

Table 1.

Parameters of input (X_1-3) and output (Y) from central composite design for optimizing extract conditions of zeaxanthin from orange paprika

Run	X1 (%, v/v)	X2 (min)	X3 (°C)	Y (mg/100 g dry weight)
1	89.9	4	68	19.1 ± 1.0
2	60.1	4	68	1.17 ± 0.05
3	75	6	100	9.95 ± 1.64
4	75	6	80	6.57 ± 0.11
5	89.9	8	68	22.0 ± 1.3
6	60.1	4	92	1.04 ± 0.04
7	75	6	80	8.03 ± 0.69
8	50	6	80	0.25 ± 0.01
9	60.1	8	92	1.31 ± 0.06
10	75	6	80	7.69 ± 1.33
11	60.1	8	68	1.35 ± 0.15
12	75	6	80	7.74 ± 0.95
13	75	6	60	7.25 ± 0.84
14	89.9	4	92	33.9 ± 0.6
15	100	6	80	24.2 ± 2.5
16	89.9	8	92	20.4 ± 1.9
17	75	9	80	19.3 ± 1.4
18	75	3	80	14.2 ± 1.7
19	75	6	80	7.90 ± 0.77
20	75	6	80	7.50 ± 0.63

Values of zeaxanthin are means and standard deviations (n = 3). X₁, ethanol concentration. X₂, extraction time. X₃, extraction temperature. Y, zeadxanthin content

RSM and ANN-GA to optimize extract conditions of zeaxanthin in orange paprika.

RSM modeling and optimization of extraction conditions were performed using Design-Expert 13 software (Stat-Ease).

ANN-GA modeling was performed according to a previous study with some modifications (Lee et al., 2022). Briefly, in this study, the optimization and modeling of process parameters for zeaxanthin extraction were performed using deep learning and optimization toolboxes in MATLAB software (R2022b, The MathWorks, Inc., Natick, MA, USA), with the aid of the ANN and GA. To conduct the nonlinear analysis, a multilayer feed-forward neural network model was utilized for mapping input and output variables. The ANN model architecture comprised three layers: an input layer with three neurons, a hidden layer with ten neurons, and an output layer with one neuron. The Levenberg‒Marquardt algorithm was employed for ANN training. Before ANN modeling, the input data were normalized between 0 and 1, while the output data were Z-score normalized. This study partitioned the experimental dataset into three randomly selected subsets: 70% for training, 15% for validation, and 15% for testing. Tangent sigmoid and linear transfer functions were applied to the hidden and output layers. The model's performance was evaluated using the mean squared error (MSE) and regression correlation coefficient (R). The best training performance of the developed ANN model was determined by identifying the lowest MSE and highest R values.

The developed ANN model was integrated with GA for input parameter optimization to maximize the output value. A double vector represented the population type. The creation, selection, mutation, and crossover functions were performed using feasible population, stochastic uniform, adaptive feasible, and scattered, respectively. A population size of 50 was selected, and the crossover fraction was set to 0.8.

The differences between the variable response values predicted by the developed RSM and ANN-GA models and the actual experimental values were expressed as relative errors.

Results and discussion

UPLC analysis of zeaxanthin

In this study, we determined the main carotenoid compositions from orange paprika using UPLC, as shown in Fig. 1(A), and the main carotenoids were confirmed to be zeaxanthin (retention time, RT: 13.4 min), lutein (RT: 14.2 min), α-cryptoxanthin (RT: 20.1 min), β-cryptoxanthin (RT: 20.3 min), and β-carotene (RT: 24.3 min). Among them, the zeaxanthin peak area was significantly higher than that of the other carotenoids. According to the findings reported by Kim et al. (2016), orange paprika is a rich source of zeaxanthin regardless of the shape and cultivation method of the paprika. Additionally, Kye et al. (2022) reported that the zeaxanthin content in orange paprika (303 ± 34 mg/kg, dry weight) was approximately 3 times higher than that in red paprika (93.2 ± 7.5 mg/kg, dry weight) and approximately 36 times higher than that in yellow paprika (8.54 ± 0.32 mg/kg, dry weight). Therefore, we used zeaxanthin as an index component to investigate the efficacy of the mathematically designed ASE method.

Fig. 1.

UPLC Chromatogram of ethanolic accelerated solvent extract from orange paprika A and zeaxanthin standards A

CCD for zeaxanthin extraction

Twenty ASE conditions were designed based on CCD (Table 1). In general, carotenoids are extracted using a nonpolar solvent such as hexane or acetone due to their nonpolar long carbon chain structure. However, these nonpolar solvent extracts have limited application in food systems and might cause health and environmental problems. Additionally, most organic solvents are inefficient for extracting carotenoids at atmospheric pressure. Therefore, we wanted to find a suitable solvent that is safe and environmentally friendly enough to be applied to food systems. Thus, this study used ethanol as an extraction solvent for extracting carotenoids; it can easily be applied to food systems and used as a GRAS solvent (Generally, Recognized as Safe, U.S. Food and Drug Administration) (Bueno et al., 2020). Under high-pressure conditions in ASE (1500 psi), ethanol can be an efficient and safe solvent for extracting carotenoids (Gunathilake et al., 2019; Kim et al., 2019). Furthermore, the carotenoid structure, characterized by a long chain of conjugated carbon double bonds, is highly susceptible to thermal degradation. Therefore, gentle extraction temperatures below 100 °C were established to prevent such degradation. The ethanol concentration, extraction time, and temperature ranges of ASE were set at 50–100%, 3–9 min, and 60–100 °C, respectively. The zeaxanthin analysis data were obtained using ASE according to CCD (Table 1) with RSM and ANN modeling.

RSM and ANN model construction

An RSM prediction model was built using experimental values obtained through CCD design, and the following regression equation was obtained:

$$\text{Y}=21.19062-0.78580{\text{3X}}_1-1.00{\text{263X}}_2-0.230{\text{116X}}_3-0.0{\text{46351X}}_1{\text{X}}_2+0.00{\text{9347X}}_1{\text{X}}_3-0.0{\text{84948X}}_2{\text{X}}_3+0.00{\text{6398X}}_1^2+0.{\text{93123X}}_2^2+0.000{\text{916X}}_3^2$$

X₁ represents ethanol concentration, X₂ represents extraction time, X₃ represents extraction temperature, and Y represents the response variable (zeaxanthin contents). The statistical parameters of the RSM model are presented in Table 2. The Model F-value of 13.53 suggests that the model holds significance. The calculated probability of observing such a large F-value solely due to noise was only 0.02%. Both terms X₁ and term X₂² are significant factors in this model (P < 0.05). On the other hand, the lack of fit F-value of 89.82 indicates a significant lack of fit (P < 0.05). The probability of obtaining such a significant lack of fit F-value due to noise alone is only 0.01%. A significant lack of fit means that the predicted RSM model was undesirable. The predicted R² value of 0.3756 deviates considerably from the adjusted R² value of 0.8558 and R² of 0.9241, exceeding the expected difference of 0.2. This result suggests a substantial block effect or a potential issue with the predicted RSM model. An observed Lack of Fit (P < 0.05) may indeed indicate potential overfitting. However, the overall equation can be meaningful if other coefficients (terms X₁ and term X₂²) remain significant (P < 0.05), and the model performance, as indicated by metrics like R² (0.9241), remains satisfactory.

Table 2.

Statistical parameters of predicted response surface methodology model for optimization of zeaxanthin extraction conditions

Source	Sum of squares	df	Mean square	F-value	p-value
Model	1513.87	9	168.21	13.53	0.0002
A-Concentration	1253.55	1	1253.55	100.84	< 0.0001
B-Time	0.5	1	0.5	0.0402	0.8451
C-Temparature	22.67	1	22.67	1.82	0.2066
AB	15.26	1	15.26	1.23	0.2938
AC	22.34	1	22.34	1.8	0.2097
BC	33.25	1	33.25	2.67	0.133
A2	29.1	1	29.1	2.34	0.157
B2	142.57	1	142.57	11.47	0.0069
C2	0.2455	1	0.2455	0.0197	0.891
Residual	124.31	10	12.43
Lack of Fit	122.94	5	24.59	89.82	< 0.0001
Pure Error	1.37	5	0.2737

The ANN model structure used in this study is shown in Fig. 2. To prevent overfitting, the ANN model was configured with only one hidden layer, as previous research has indicated that utilizing more than two hidden layers could potentially result in overfitting (Shao et al., 2007). In addition, accurately determining the number of neurons in the hidden layers is crucial for developing an optimal model capable of accurately predicting the output. Too few neurons in the hidden layers can lead to poor performance, while too many neurons in the hidden layers can lead to overfitting (Xi et al., 2013). Hence, the ideal number of neurons in the hidden layer was determined by identifying the minimum MSE value between the predicted and experimental data. This analysis revealed that the optimal number of neurons in the hidden layer was ten. After four iterations, the training process of the ANN model was completed, achieving the lowest MSE. The weights and bias values of the trained ANN utilized in this study can be found in Eqs. (1–4).

$$U=\left[\begin{array}{ccc}1.2469& 1.5848& -2.2480\\ 0.7583& -0.8719& 2.8519\\ 0.3053& 1.6301& 2.3052\\ -1.9143& 2.1090& 0.5146\\ 1.6011& -2.0310& 1.6473\\ 2.5955& 1.5594& -0.4216\\ 1.2547& 2.6574& -1.0121\\ -2.3386& -0.9081& 1.5954\\ -1.6967& 1.9951& 1.2361\\ 0.4630& 2.7936& 0.1485\end{array}\right]$$ 1

$$TH=\left[\begin{array}{c}-3.0094\\ -2.2981\\ -1.7073\\ 1.2004\\ -0.2712\\ 0.2541\\ 0.9213\\ -1.7427\\ -2.5475\\ 3.2027\end{array}\right]$$ 2

$$W=\left[0.3205-0.39800.1371-0.3529-0.12931.0835-0.6595-0.04350.2285-0.3034\right]$$ 3

$$TO=\left[0.4383\right]$$ 4

Fig. 2.

Architecture of artificial neural network model used in this study

The connection weight and bias values between the input and hidden neurons are represented as U and TH, respectively. The connection weight and bias values between the hidden and output neurons are represented as W and TO, respectively.

Figure 3(A) shows the performance of MSE as the number of epochs increases during training, validation, and testing, and it demonstrates that the network learns effectively as the number of epochs increases. The network performance of the optimized topology was measured during development using the metrics requiring the fewest training and testing errors. The number of epochs was kept to a minimum to establish the optimal topology. This approach was adopted because higher epochs can potentially result in overfitting issues with the model (Shao et al., 2007). It limits the number of epochs aimed to balance training the model effectively and avoiding overfitting. The Levenberg‒Marquardt algorithm was used for network training, and the best validation performance was founded on ethanol concentration (X₁), extraction time (X₂), and extraction temperature (X₃). This study obtained the best validation performance at Epoch 2, with an MSE value of 0.042236 (Fig. 3A).

Fig. 3.

Network training curves with epochs number for trained subsets A and regression analysis of the generated artificial neural network model B in this study. R, correlation coefficient

Upon minimizing the MSE after six iterations, the ANN model's training was deemed complete. A regression analysis (Fig. 3B) was conducted to assess the fitness of the predicted and actual values of zeaxanthin content derived from 20 experiments. The R values for the training, validation, testing, and all datasets were found to be 0.99954, 0.99982, 0.99946, and 0.99451, respectively, as depicted in Fig. 3(B). The MSEs were 0.0010, 0.0422, 0.0429, and 0.0135 for the training, validation, testing, and all datasets, respectively. When the MSE is close to 0, this value indicates a good fit of the model, meaning that the predicted values closely match the experimental values. A high R, close to 1, also signifies a strong relationship between the experimental and predicted values (Lee et al., 2022). Therefore, these results demonstrated that the development of the ANN model was successful (Table 3).

Table 3.

Predicted and experimental values at optimal zeaxanthin extraction conditions from orange paprika using RSM and ANN-GA

	X1 (%, v/v)	X2 (min)	X3 (°C)	Y (mg/100 g)	Relative error (%)
Predicted value (RSM)	99.9	3.1	99.6	50.7	10.46
Predicted value (ANN-GA)	100	3.4	99.2	44.9	2.18
Experimental value	100	3	100	45.9 ± 1.2*

Predicted values were obtained by optimization using response surface methodology and artificial neural network coupled with genetic algorithm. X₁, ethanol concentration. X₂, extraction time. X₃, extraction temperature. Y, zeaxanthin content. ^*Value is mean and standard deviation (n = 3)

Effects of ASE conditions on zeaxanthin extraction

The nonlinear relationships between the ASE conditions used as input and the zeaxanthin content obtained as output were graphically represented using the RSM model (Fig. 4) and computation results obtained from the trained ANN model (Fig. 5). In comparison to extraction time and temperature, the concentration of ethanol used as the extraction solvent had the most significant impact on zeaxanthin extraction from orange paprika. The zeaxanthin content was the highest when the ethanol concentration was 89.9% (Table 1). However, the zeaxanthin content tended to decline drastically when the ethanol concentration was lower than 89.9% (Fig. 4). Koo et al. (2012) reported that ethanol was the best solvent for zeaxanthin extraction under pressurized liquid extraction (1500 psi) from Chlorella ellipsoidea, and the extraction yield of zeaxanthin was sensitively affected by the extraction time. However, extraction time was not a significant factor in the extraction yield in this study. Wang et al. (2019) report that ethanol was the most effective extraction solvent with strong cell-penetrating ability, enabling rapid extraction of zeaxanthin from samples. They also observed that the ethanol concentration significantly impacted the zeaxanthin content, regardless of the extraction temperature and time. When the zeaxanthin content was compared between experimental runs 1 and 2 (or 7 and 8) in Table 1, the zeaxanthin content rapidly changed with different ethanol concentrations at the same extraction time and temperature. These results suggest that the factor most affecting the extraction of zeaxanthin from orange paprika was the ethanol content, which may be due to the nonpolar structure of zeaxanthin.

Fig. 4.

Three-dimensional plots of the relationships between input (ethanol concentration, extraction time, and extraction temperature) and output (zeaxanthin) parameters of response surface methodology in this study

Fig. 5.

Three-dimensional plots of the relationships between input (ethanol concentration, extraction time, and extraction temperature) and output (zeaxanthin) parameters of artificial neural network coupled with genetic algorithm in this study

Optimization of extraction conditions with RSM or GA and zeaxanthin extracted under optimal conditions

The ASE conditions (ethanol concentration, extraction time, and extraction temperature) were optimized for the maximum output parameter (zeaxanthin content) using the RSM and GA after the ANN model had been fully developed (Table 2). The expected optimal extraction conditions of RSM (99.9% ethanol, 3.1 min, and 99.6 °C) and ANN-GA (100% ethanol, 3.4 min, and 99.2 °C) were nearly identical; however, there was a difference in the predicted response values as zeaxanthin content, with RSM predicting 50.7 mg/100 g dry basis and ANN-GA predicting 44.9 mg/100 g dry basis. The actual extraction process was conducted after setting the ASE parameters as close as possible to the optimal extraction conditions predicted by each model. Under the extraction conditions, the actual experimental value (zeaxanthin content) was 45.9 ± 1.2 mg/100 g dry weight, which was similar to the predicted value of the highest zeaxanthin content under ANN-GA, 44.9 mg/100 g dry weight. The relative error between the experimental and predicted response values under the ANN-GA model was 2.18%; however, the relative error under the RSM model was 10.46%. These findings show that the ANN with GA used in this study had more excellent prediction and optimization capabilities than RSM.

Under optimal extraction conditions of zeaxanthin using ASE, the orange paprika contained 45.9 ± 1.2 mg/100 g dry weight of zeaxanthin. The concentration of zeaxanthin obtained from orange paprika may be adequate to produce bioactive outcomes. Compared to other carotenoids, zeaxanthin and lutein play a crucial role in maintaining the health of the eyes and brain. These two carotenoids constitute 66–77% of the total carotenoid population in the brain. Their presence is significant for preserving the well-being of the eyes and brain (Krinsky & Johnson, 2005). For three months, patients with diabetic retinopathy received 0.5 mg/d zeaxanthin and 6 mg/d lutein, and the treatments reduced diabetic macular edema and increased contrast sensitivity (Hu et al., 2011). Zeaxanthin (oral dose of 20 mg/kg) slowed the development of obesity and improved dyslipidemia in mice with obesity induced by a high-fat diet (Liu et al., 2017). Moreover, in a study conducted by Tuzcu et al. (2017), it was found that the administration of marigold flower extract containing 13.5 mg/kg of zeaxanthin through gavage resulted in a decrease in malondialdehyde production induced by a high-fat diet, and also restored the total antioxidant capacity in the retina of rats. Additionally, 25 mg/kg zeaxanthin administration to Mongolian gerbils brought fibrosis and lipid peroxidation back to baseline, indicating that this carotenoid may help treat nonalcoholic steatohepatitis (Chamberlain et al., 2009). Therefore, there may be enough zeaxanthin in orange paprika as a bioactive source.

In this study, we have confirmed that the optimization of extraction methods using ANN-GA is more efficient than RSM. We also determined the number of hidden layers and neurons for ANN-GA modeling by referencing prior research to mitigate overfitting (Lee et al., 2022). However, despite these efforts, the experiment design involving 20 trials based on the traditional 3-factor RSM approach may still lead to overfitting in ANN-GA modeling. Therefore, in future research, a more accurate ANN-GA modeling approach for preventing overfitting and optimization prediction will be required. In response to this concern, we are considering increasing the number of experiments. We plan to collect and analyze more experimental data to identify the optimal size of the hidden layer, with a strong emphasis on enhancing model performance. Through these efforts, we anticipate finding effective measures to address overfitting issues and improving the reliability of the model.

In conclusion, this study investigated optimizing ASE conditions for zeaxanthin from orange paprika. Among the detected carotenoids, zeaxanthin was the highest in orange paprika. The experimental results in this study indicated that the ANN model exhibited accurate predictive capabilities, as evidenced by high R values and low MSE values. This outcome signifies a strong agreement between the predicted and actual values, indicating a higher model generalization level than the RSM model. The developed model in this study is a correct and efficient tool for predicting and optimizing the extraction conditions of maximum zeaxanthin from orange paprika. These results indicated that hydrophobic zeaxanthin could be recovered using ethanol, a green solvent, using ASE and that ANN-GA modeling might increase extraction efficiency. Therefore, combining ASE and ANN-GA modeling might benefit the effective and environmentally friendly extraction of hydrophobic functional elements from food resources.

Supplementary Information

Below is the link to the electronic supplementary material.

Data availability

The data that support the findings of this study are available on request from the corresponding author.

Declarations

Conflict of interest

The authors declare that they have no conflict of interest.